Timestamps

Introduction

The aim of the timestamp library is to make it easier for different boards to save timestamps in cbmem / stash (until cbmem is brought up) by providing a simple API to initialize, add and sync timestamps. In order to make the timestamps persistent and accessible from the kernel, we need to ensure that all the saved timestamps end up in cbmem under the CBMEM_ID_TIMESTAMP tag. However, until the cbmem area is available, the timestamps can be saved to a SoC-defined _timestamp region or in a local stage-specific stash. The work of identifying the right location for storing timestamps is done by the library and is not exposed to the user.

Working of timestamp library from a user perspective can be outlined in the following steps:

Initialize the base time and reset cbmem timestamp area
Start adding timestamps

Behind the scenes, the timestamp library takes care of:

Identifying the correct location for storing timestamps (cbmem or timestamp region or local stash).
Once cbmem is up, ensure that all timestamps are synced from timestamp region or local stash into the cbmem area.
Add a new cbmem timestamp area based on whether a reset of the cbmem timestamp region is required or not.

Transition from cache to cbmem

To move timestamps from the cache to cbmem (and initialize the cbmem area in the first place), we use the CBMEM_INIT_HOOK infrastructure of coreboot.

When cbmem is initialized, the hook is called, which creates the area, copies all timestamps to cbmem and disables the cache.

After such a transition, timestamp_init() must not be run again.

Data structures used

The main structure that maintains information about the timestamp cache is:

struct __packed timestamp_cache {
        uint16_t cache_state;
        struct timestamp_table table;
        struct timestamp_entry entries[MAX_TIMESTAMP_CACHE];
};

cache_state

The state of the cache is maintained by cache_state attribute which can be any one of the following:

enum {
        TIMESTAMP_CACHE_UNINITIALIZED = 0,
        TIMESTAMP_CACHE_INITIALIZED,
        TIMESTAMP_CACHE_NOT_NEEDED,
};

By default, if the cache is stored in local stash (bss area), then it will be reset to uninitialized state. However, if the cache is stored in timestamp region, then it might have garbage in any of the attributes. Thus, if the timestamp region is being used by any board, it is initialized to default values by the library.

Once the cache is initialized, its state is set to CACHE_INITIALIZED. Henceforth, the calls to cache i.e. timestamp_add know that the state reflected is valid and timestamps can be directly saved in the cache.

Once the cbmem area is up (i.e. call to timestamp_sync_cache_to_cbmem), we do not need to store the timestamps in local stash / timestamp area anymore. Thus, the cache state is set to CACHE_NOT_NEEDED, which allows timestamp_add to store all timestamps directly into the cbmem area.

table

This field is represented by a structure which provides overall information about the entries in the timestamp area:

struct timestamp_table {
        uint64_t        base_time;
        uint32_t        max_entries;
        uint32_t        num_entries;
        struct timestamp_entry entries[0]; /* Variable number of entries */
} __packed;

It indicates the base time for all timestamp entries, maximum number of entries that can be stored, total number of entries that currently exist and an entry structure to hold variable number of entries.

entries

This field holds the details of each timestamp entry, up to a maximum of MAX_TIMESTAMP_CACHE which is defined as 16 entries. Each entry is defined by:

struct timestamp_entry {
        uint32_t        entry_id;
        uint64_t        entry_stamp;
} __packed;

entry_id holds the timestamp id corresponding to this entry and entry_stamp holds the actual timestamp.

For timestamps stored in the cbmem area, a timestamp_table is allocated with space for MAX_TIMESTAMPS equal to 30. Thus, the cbmem area holds base_time, max_entries (which is 30), current number of entries and the actual entries represented by timestamp_entry.

Function APIs

timestamp_init

This function initializes the timestamp cache and should be run as early as possible. On platforms with SRAM, this might mean in bootblock, on x86 with its CAR backed memory in romstage, this means romstage before memory init.

timestamp_add

This function accepts from user a timestamp id and time to record in the timestamp table. It stores the entry in the appropriate table in cbmem or _timestamp region or local stash.

timestamp_add_now

This function calls timestamp_add with user-provided id and current time.

Use / Test Cases

The following cases have been considered while designing the timestamp library. It is important to ensure that any changes made to this library satisfy each of the following use cases:

Case 1: Timestamp Region Exists (Fresh Boot / Resume)

In this case, the library needs to call timestamp_init as early as possible to enable the timestamp cache. Once cbmem is available, the values will be transferred automatically.

All regions are automatically reset on initialization.

Case 2: No timestamp region, fresh boot, cbmem_initialize called after timestamp_init

timestamp_init will set up a local cache. cbmem must be initialized before that cache vanishes - as happens when jumping to the next stage.

Case 3: No timestamp region, fresh boot, cbmem_initialize called before timestamp_init

This case is not supported right now, just don’t call timestamp_init after cbmem_initialize. (Patches to make this more robust are welcome.)

Case 4: No timestamp region, resume, cbmem_initialize called after timestamp_init

We always reset the cbmem region before using it, so pre-suspend timestamps will be gone.

Case 5: No timestamp region, resume, cbmem_initialize called before timestamp_init

We always reset the cbmem region before using it, so pre-suspend timestamps will be gone.